Amazon Textract Insights extracts fields from mixed document packets

Amazon has introduced Textract Insights, a managed service that pulls fields from scanned documents and mixed packets using plain-language descriptions. According to Amazon, every extracted value comes with a confidence score and a reference to the page it was taken from.
What the service does
Users set out the information they need in a blueprint, a JSON Schema in which each field carries a description of what to look for. Each field is also marked as either read word for word from the page or inferred from the document's content. Amazon states that no model training is needed.
Amazon names AWS public blueprints as a starting point, covering common documents such as a W-2, Form 1040, Form 1098, a passport and a birth certificate. Organisations can also write blueprints for their own paperwork.
Handling mixed packets
Amazon presents mixed packets, such as claims files and onboarding paperwork that combine typed forms, dense tables, scanned receipts and free text in no fixed order, as the central difficulty. Under the new approach, a single file of up to 3,000 pages is submitted from an Amazon S3 bucket with splitting switched on. The service breaks it into logical sub-documents, assigns each to the matching blueprint and returns the extracted fields.
As an example, Amazon cites a mortgage file: one 26-page PDF containing a W-2, payslips, bank statements and handwritten notes from a loan officer. Once the project is set up, processing it takes one StartDocumentInsights call.
Alternatives and their limits
Amazon sets the service against two common approaches. A custom pipeline of optical character recognition, splitting, classification, a foundation-model prompt and a review step has to be built and operated in-house. A general-purpose large language model gives fluent answers but, Amazon says, does not reliably name its source page or provide a confidence score, which makes it hard to tell which values can be accepted automatically.
Open questions
- How closely does the confidence score track actual extraction errors on real documents?
- What does the service cost at the 3,000-page limit?